Urban Growth Derived from Landsat Time Series Using Harmonic Analysis: A Case Study in South England with High Levels of Cloud Cover

Round 1

Reviewer 1 Report

This study developed a method that identifies change in the urban fabric at the pixel level based on breaks in the seasonal and year-on-year trend of the Normalised Difference Vegetation Index (NDVI). The method proved relatively robust for outliers and missing data, for example in the case of high levels of cloud cover, but does rely on a period of data availability before and after the change event. The corresponding overall accuracy is 91.8%, the user’s accuracy is 50.9% and the producer’s accuracy is 69.0%. This study is suitable for publishing in Remote Sensing. However, the manuscript has some major shortcomings:

 

(1) The overall accuracy is wrongly calculated. The user’s accuracy is 50.9% and the producer’s accuracy is 69.0%, so the overall accuracy is ~60%. You should not include the samples without any detected changes in the calculation of overall accuracy. For example, if there are 10000 samples and only one changed sample, and your method can not detect the change, the overall accuracy using your calculation is > 99%. This does not make sense. You should refer to the validation of the COLD algorithm (Zhu et. al., 2020).

 

(2) There is no harmonization of reflectance difference between Landsat 5, 7, and 8. The reflectance difference between Landsat 7 and 8 can reduce the detection accuracy, you should first reduce the reflectance difference between different Landsat sensors.

 

(3) Your method can only detect one change event during the long time period between 1984 and 2018. This can not be accepted. The multiple change events during this time period can be commonly observed.

 

(4) Many changes can not be detected by only using NDVI, and some other vegetation indexes should be integrated.

 

Therefore, I cannot recommend publishing in the current form.

Author Response

Dear Reviewer,

We would firstly like to thank you for your constructive comments which have proved valuable to the improvement of this paper.  We would like to particularly thank you for raising comment 1, and we hope our further clarifications have addressed this issue, and all others raised. 

We have included below your original comments, our responses, followed by the action we have taken in the manuscript to address your concerns. 

We hope that you find these explanations satisfactory. 

 

(1)    The overall accuracy is wrongly calculated. The user’s accuracy is 50.9% and the producer’s accuracy is 69.0%, so the overall accuracy is ~60%. You should not include the samples without any detected changes in the calculation of overall accuracy. For example, if there are 10000 samples and only one changed sample, and your method can not detect the change, the overall accuracy using your calculation is > 99%. This does not make sense. You should refer to the validation of the COLD algorithm (Zhu et. al., 2020). 

 

Thank you for requesting clarification on the measure of accuracy.  In the paper, ground truth data is classified in three types of change: change, partial-change, and no-change, however the model only differentiates between change and no-change.  This results in a 3x2 confusion matrix.  In the partial-change class the majority of the pixel remains unchanged. We treat this class as no change where a 2x2 confusion matrix is necessary.  Therefore, for the calculation of user’s accuracy, overall accuracy, Kappa, and F1 score, partial-change is reclassed as no-change.  However, for the producer’s accuracy this was not necessary and separate statistics are calculated for each of the classes of change, partial-change, and no-change.  The other statistic to make full use of the 3x2 confusion matrix is Weighted Kappa, which assumes full agreement between partial-change and no-change, and partial agreement between partial-change and change. 

The use of a 3x2 matrix and accounting for partial agreement, makes the accuracy assessment slightly unusual, whilst providing more insight on the nature of disagreement. Furthermore, this method more accurately details the total level of land cover of change, where in more traditional analysis, partial-change would not have been considered as having undergone change. 

For the sake of clarity, we have stressed these points in the sections below:

 

Section 3.4.3 (lines 305-309):

“conversely, producer’s accuracy (PA) was calculated separately for the no-change and partial-change classes, to provide better insights into the nature of disagreements. This results in a 3x2 confusion matrix, rather than the more conventional tables 2x2 in a binary classification.”

Figure 4, caption:

“Note that for UA, OA, K, and F1 score, partial-change is counted as no-change..  For WK, partial-change is in half agreement with change, and full agreement with no-change.”

Table 4, caption:

“Note that for UA, OA, K, and F1 score, partial-change is counted as no-change..  For WK, partial-change is in half agreement with change, and full agreement with no-change.  “

 

 

Thank you also for drawing our attention to the validation of the COLD algorithm, we have now added the F1 score as an additional measure of accuracy.  The value was calculated for all thresholds tested, the validation period, the time of change analysis, and also for the PCC. 

The method has been added to section 3.4.3 (lines 313 – 315):

“F1 score (or F measure) was calculated to quantify the balance between producer’s and user’s accuracy of the change class (Zhu et al., 2020).”

F1 scores have been added to Figure 4 and Table 4. 

Results have been added on lines 384 - 386

“F1 score is approximately equal for all thresholds between 0.88 and 0.93, with values reducing outside this range.”

And lines 440 – 441:

“The great disparity between user’s and producer’s accuracy of change is reflected by the low F1 score (Table 4).”

And lines: 580 – 582:

“The disparity between the PA and UA of the PCC is clearly demonstrated via the lowest F1 score of any change detection.  The greatly increased PA comes at the expense of a reduced UA, and is reflected in other statistics, particularly the K.  “

 

(2)    There is no harmonization of reflectance difference between Landsat 5, 7, and 8. The reflectance difference between Landsat 7 and 8 can reduce the detection accuracy, you should first reduce the reflectance difference between different Landsat sensors.

Thank you for raising this point, which should have been clarified in the manuscript.

There are indeed differences among NDVI values derived from different Landsat sensors.  These do require harmonization when absolute vegetation is to be quantified and compared among different dates, and when inter and intra-annual index variations are needed to inform of crop type, health, etc. In our case, however, the time series of NDVI is only needed to estimate long time average dynamics, with the goal to discriminate between those typical of natural and farmed vegetation from those of an urban environment.  A single harmonic trend is fitted to natural and farmed pixels regardless their vegetation type or specific year performance, both with higher impact on the NDVI than the use of different Landsat sensors’ data. 

At the beginning of our work, we retrieved and investigated many of these Landsat 5, 7 and 8 combined NDVI time series and made sure that the mild mismatch between sensors’ indices didn’t have a noticeable impact on the retrieved long-time average trends (NDVI linear trend and amplitude). We have clarified this point in the manuscript in lines 139 – 142 , included as follows:

“In this study, no harmonisation between Landsat sensors was performed as during preliminary analysis, the NDVI calculated from Surface Reflectance images was observed to have negligible impact on long term average trends (the NDVI linear trend and amplitude, calculated below).  “

(3)    Your method can only detect one change event during the long time period between 1984 and 2018. This can not be accepted. The multiple change events during this time period can be commonly observed. 

The method is indeed only able to identify (at most) one change event over the analysed time period. For many change processes and landscape types this will not be reasonable assumption, but in our case the type of change that we are interested in is the transition from non-urban to urban land. This type of change is typically one-directional and permanent, and for that reason it is a reasonable assumption here. We have now explained this as a limitation in the discussion (lines 620-627) and future directions of this work could include the identification of multiple changes within the analysed time period.  Our method also outperforms a post classification comparison, which likewise can only detect a single change. 

Section 5.2 has now been expanded to improve the discussion of this assumption:

“Urban growth is typically one-directional, therefore multiple changes are unlikely.  The purpose of our study was to detect this type of change, therefore the detection of a single change is a reasonable assumption, however may not be universally true.  Only a single pixel in all those analysed underwent two land cover changes in the period between 2002-2015 (excluding transitionary land cover types such as worksites), however partial-change­ was often associated with longer term incremental changes totalling < 50% of the pixel (such as extensions and garden development over several years).”

 

(4)    Many changes can not be detected by only using NDVI, and some other vegetation indexes should be integrated. 

As previously mentioned, our main goal is to detect land conversion to urban. Our harmonic approach applied only to NDVI has already outperformed the post-classification change detection, however, we completely agree with the reviewer that the method’s accuracy could be improved by the use of other indices, e.g. the build-up index. Furthermore, other indices could yield improvements in classification accuracy and allow interpretation of vegetation to vegetation changes.  

We have included several references to provide a more solid foundation for the sole selection of NDVI as a change indicator in section 3.4.1 (lines 235 – 238):

“NDVI was chosen as the single metric for change detection as a vegetation index it is subject to periodic cycles and a is a (counter) indication of urbanization. Other studies have successfully detected land cover changes using solely NDVI and derived statistics [11], [38], particularly sinusoidal function change detection methods [18], [39].”We have added the lines (605 – 607) below to stress the fact that our approach can be applied to time series of other indices too:

“To improve change detection and classification accuracy, the inclusion of other vegetation indices is an avenue for further research. This can be achieved by the substitution of NDVI for other indices into equations 1, 2, and 3.”

 

 

We would like to offer thanks once again to the referee for their time and comments. 

 

 

 

 

Author Response File: Author Response.docx

Reviewer 2 Report

This study investigated the use of structural break detection in harmonic analysis to detect and classify land cover change in the context of urban growth. They developed a method that can identify the changes in urban fabric at the pixel level based on breaks from the NDVI time series data. The proposed method performs well in regions where changes exhibit significant differences of NDVI dynamics amongst land cover types and less so where minor differences of NDVI are observed. I suggest a major revision of this paper after addressing some critical bullets in this study. Below are detailed comments.

 

Main concerns:

1. This paper begins to deny the common method of land cover change research: post-classification comparison (PPC), then a new method based on NDVI is proposed, but there is no comparison between the two precision in the same study area and the same time series. I suggest to add the comparison of these two approaches.
2. According to the annual analysis in this paper, it is unclear about the pixels used for composition. Are all pixels without cloud contamination were used, or only pixels those within a specific region with less cloud coverage were used? Please clarify it.
3. In terms of data selection, although Landsat data are available with a long historical archive, its spatial resolution are relatively coarse (30 m pixels). Therefore, the Landsat data cannot accurately reflect many changes in urban landscapes. Is Sentinel2 data can be used for verification? The accuracy of the proposed method under different resolutions should be discussed.
4. Many of the classifications in this article are based on manual classifications, which are subjective at a certain extent. Please discuss about it.

 

Minor comments:

1. Figure 1:it is better to add local enlarged view of study area
2. Table 2: pleased add manually classified criteria
3. Lin 201: why the second stage uses a Random Forest method to classify the type of change that occurred in the change pixels.
4. What are the physical meanings of the four coefficients of A, B, C, D constructed in 3.4.2
5. Is the distribution of random sampling points balanced in space?
6. Lin 321: what the critical area?

Author Response

Dear Reviewer,

 

Thank you for your time to offer us constructive comments of our manuscript.  We particularly appreciate your desire for discussion on the impact of changing resolution on this study, and hope you find our discussion satisfactory.  We have addressed all of your comments in our revised manuscript. 

We have included below your original comments, our responses, followed by the action we have taken in the manuscript to address your concerns. 

We hope that you find these explanations satisfactory. 

 

(1)    This paper begins to deny the common method of land cover change research: post-classification comparison (PPC), then a new method based on NDVI is proposed, but there is no comparison between the two precision in the same study area and the same time series. I suggest to add the comparison of these two approaches.

 

Actually, we included this comparison in Section 4.2. We apologise if this comparison was not clearly presented.  The LCM maps are nominally titled “2007” and “2015” however they were created using images from a range of dates, including data from the end of 2005.  To facilitate comparison, our validation trial uses all Landsat images from 2006 to 2015 to better coincide with the data used, rather than the names of the dataset.  This has been clarified as follows:

Section 3.3 (lines 213 – 215):

“LCM 2007 uses images ranging from 02/09/2005 to 18/07/2008, and LCM 2015 images from 01/01/2014 to 10/12/2015  [35], [37].  “

Section 3.4.3 (lines 298 – 300):

“to coincide with the images used to produce the LCM 2007 and 2015, rather than their nominal dates.”

 

(2)    According to the annual analysis in this paper, it is unclear about the pixels used for composition. Are all pixels without cloud contamination were used, or only pixels those within a specific region with less cloud coverage were used? Please clarify it.

 

Thank you for raising this query. Only individual pixels identified as being affected by cloud or cloud shadow by the CFMask algorithm were excluded from the analysis.  This was clarified in section 3.1 (lines 132 – 135):

“Any pixel in any image which was identified by the CFMask as being contaminated with clouds, shadow, water, or snow, was removed from the analysis, however corresponding pixels in other images remained in the analysis if they were identified as cloud-free. “

(3)    In terms of data selection, although Landsat data are available with a long historical archive, its spatial resolution are relatively coarse (30 m pixels). Therefore, the Landsat data cannot accurately reflect many changes in urban landscapes. Is Sentinel2 data can be used for verification? The accuracy of the proposed method under different resolutions should be discussed.

We acknowledge the spatial resolution limitations of Landsat raised by the reviewer, this mission is one of the few satellite data sources that enables a historical reconstruction of urban growth from the 80’s. 

For much of this period, Landsat imagery represented the highest spatial resolution available, and indeed is still used for the basis of land cover products and the detection and quantification of change.  Sentinel unfortunately does not cover the time period analysed.  Quickbird images represent a high-resolution archive, however we deemed the higher resolution of Google Earth imagery and widespread accessibility to be preferable (lines 145 – 147):

“Google Earth images represent the highest resolution source of ground truth data available to this study,”

 

We have included in the discussion a new section (lines 653 – 664):

“5.4 Impact of changing spatial resolution

The 30 m resolution of Landsat is well suited to the detection of housing unit construction but fails to adequately capture finer resolution changes such as small increases of paved surface in gardens.  Increasing the spatial resolution of the sensor (such as using Sentinel 2) should not impact the accuracy of change detection in areas where the change is larger than the Landsat pixel size, but will aid the detection of smaller changes, that would be partial-change at the Landsat resolution but complete change at the Sentinel 2 one.

Theoretically, this method requires only a single year of time series data before and after the change detection period. But it is expected to be more accurate and advantageous than conventional pair date comparisons for longer periods, such as that of the Landsat archive. However, the ideal time series length is subject to further research.” 

 

(4)    Many of the classifications in this article are based on manual classifications, which are subjective at a certain extent. Please discuss about it.

 

The manual classification is based on high-resolution Google Earth images.  The manual classification includes the occurrence of change (no-change, partial change, change) and the land cover before and after the change (vegetated, vegetated with minor structures, bare ground, sparsely built-up, and densely built-up). The classification is ultimately subjective but based on readily identifiable features in the images. We have high confidence in the manual classification and have assumed it accurate in the analysis. The supplementary data includes details of all manual classifications, a description of the observed change, and the dates at which the change occurs to allow reproduction and scrutiny.  

We have included in section 3.2 the following (lines 166 – 171):

“and a short description of the change occurring.  The five land cover classes were chosen to be relevant to urbanisation and readily identifiable in high-resolution images.  The two major classes, urban and vegetation, used in the classification of change, are aggregations of sparsely built-up, densely built up, and bare surface; and vegetation, and vegetated with minor structures respectively. The pixel sets used were:”

 

And lines 198 – 200

“(Olofsson et al., 2014) discusses issues with the manual classification of ground truth pixels.  To address these issues, all the classifications, timings, and justifications are included in the supplementary data (S1 – Points)”

 

 

Figure 1:it is better to add local enlarged view of study area

We have added the study area to Figure 1

Table 2: pleased add manually classified criteria

We have clarified the defining characteristics as the classification criteria.  Such broad classes were used to increase the robustness of classifications and are readily discernible in high-resolution imagery.  

Land cover class	Abbreviation in text	Defining characteristics and classification criteria	Major Class	Major class abbreviation in text
Sparsely built-up	SBU	A mix of buildings and vegetation	Urban	U
Densely built-up	DBU	Dominated by buildings, vegetation largely absent	Urban	U
Bare ground	BG	No buildings or vegetation (e.g. car parks, bare soil)	Urban	U
Vegetated	VE	Farmland or grass	Vegetation	V
Vegetated with minor structures	VMS	Vegetated land with presence of small structures, paths, water, shrubs, or trees	Vegetation	V

 

 

Lin 201: why the second stage uses a Random Forest method to classify the type of change that occurred in the change pixels.

We have added justification to the random forest classifier on lines 243 – 245: “The supervised Random Forest classifier was selected due to its non-parametric nature, potential for high accuracy results (Rodriguez-Galiano et al., 2012) and wide use within GEE based studies (e.g. (Gumma et al., 2020; Huang et al., 2017; Oliphant et al., 2019)).”

 

What are the physical meanings of the four coefficients of A, B, C, D constructed in 3.4.2

The physical meanings of the coefficients have been added to section 3.4.5 (lines 259 – 261): “. Where, cis the slope of the linear trend of mean NDVI and d its intercept. The parameters a and b describe the oscillation around the mean and are more readily interpreted when transformed”  

 

Is the distribution of random sampling points balanced in space?

We have added discussed the spatial distribution of the sample points in section 3.2 (lines 193 – 197): “Sets A and B were randomly selected from across Swindon and are therefore spatially balanced.  Sets C and D were randomly selected from those pixels identified as change in the validation map (see below).  Set E was oversampled in an area of substantial urban growth to test the methods dating capability and is therefore not spatially balanced.”

 

Lin 321: what the critical area?

Thank you for spotting this, it has been deleted

 

We would once again like to add our thanks to the reviewer for their valuable comments and hope we have satisfactorily addressed your concerns. 

 

 

 

Author Response File: Author Response.docx

Reviewer 3 Report

The ms is written quite well. I have a few comments / questions for authors to address before the ms can be considered for publication.

Major comments:

Since the authors explicitly aim to characterize urbanization in area with high-level of cloud. In my opinion, the authors should demonstrate how number of observations (level of cloudiness) influences performance of the proposed method (to see how far this method can push).

Discussion should discuss a bit about computational aspect of the proposed method as well. I believe it would take quite long to process a large study area (especially when higher resolution imagery is used, e.g., Seninel-2). I also wonder if we use linear trends (or simple mean NDVI) instead of harmonic curves to detect time of change to reduce computational time of this process.

Minor comments:

1. Lines 40-46: Incomplete literature review. How about PCC based on a multiple-year dataset (Lark et al.,2015; Nguyen et al., 2019) or a ground sample dataset (Olofsson et al., 2014)?

Lark, T.J.; Salmon, J.M.; Gibbs, H.K. Cropland expansion outpaces agricultural and biofuel policies in the United States. Environ. Res. Lett. 2015, 10, 044003.

Olofsson, P., Foody, G.M., Herold, M., Stehman, S. V., Woodcock, C.E., Wulder, M. a., 2014. 745 Good practices for estimating area and assessing accuracy of land change. Remote Sens. Environ. 148, 42–57.

2. Lines 71-72: This assumption may not be true for many regions around the world. Why makes this assumption?
3. Figure 1: texts on the insert map are not clear, remove coordinates of the insert map to avoid confusion.
4. Line 117: remove “atmospherically corrected”
5. Lines 123-125: do the authors do anything to mitigate this problem? If not, don’t mention it here.
6. Section 3.2 and 3.3: It would be perfect to couple the ground dataset with the UK’s land cover maps using the protocol presented in Olofsson et al., 2014.
7. Line 240: h should be between 0.85 and 1 (as implemented in Figure 4)
8. Section 3.4.5: should state all classified classes and their abbreviations here.
9. Lines 292-293: should put “V – V”, “V – U”, etc... here
10. Figure 8: set up abbreviations before using them.
11. Table 4, 8: can use abbreviations to shorten row/column names (e.g., OA, PA, UA, V -V)

Author Response

Dear Reviewer,

 

We would like to extend our thanks for your time and effort you committed to providing us with a constructive commentary of our manuscript.  We particularly appreciate your desire for discussion of the impact of clouds on our algorithm and hope you find the response we provide to be satisfactory. 

We have included below your original comments, our responses, followed by the action we have taken in the manuscript to address your concerns. 

We hope that you find these explanations satisfactory. 

 

Since the authors explicitly aim to characterize urbanization in area with high-level of cloud. In my opinion, the authors should demonstrate how number of observations (level of cloudiness) influences performance of the proposed method (to see how far this method can push).

Thank you for raising this important point, which was not adequately covered in our manuscript. An additional short section has been included to address it.  For each pixel, the percentage of cloud free observations was calculated.  This was then separated into classes and a table constructed similar to a confusion matrix, but showing the average ratio.  The data shows (potentially due to the high minimum number of cloud free pixels in any location) there is no impact on accuracy.  The potential effect of further cloud cover could be investigated through simulation, i.e. by reducing the number of image available, which we believe is beyond the scope of this proof of concept. 

Two sections have been added.  In the results (lines 531 – 548):

“4.6 Number of clouds per-pixel

The number of observations per-pixel ranged between 697 and 731, as not all images included the entire study area. Per-pixel minimum and maximum cloud-free observations were: 72 (a highly mixed SBU pixel) and 343 (a pure VE pixel), respectively.  These two pixels were investigated individually and were both correctly identified as undergoing no-change. 

The cloud free coverage for all pixels in Set B was tabulated by both model and ground truth classification (Table 9). The minimum cloud free coverage for a pixel was 158 (21.9%), and the maximum 332 (45.8%).  There is no obvious correspondence between cloud free coverage and change classification.

Out of the 500 pixels of Set B, the 50 pixels with the highest number of cloud free pixels, and the lowest 50 pixels were inspected to determine any correlation to land cover class.  Of the lowest 50 pixels, only 10 involved rural classes (either change to for from V, or stable V), conversely of the highest 50 pixels, 35 involve the V class.  DBU is absent from the top 50 pixels, and VE is absent from the lowest 50 pixels.  

 

Table 9. The percentage of cloud free pixels in Set B by change classification

		Ground truth
		Change	No-change
			Partial-change	No-change
Model	Change	41.0	41.8	43.2
	No-change	41.8	42.0	42.2

"

 

 

And in the Discussion (lines 668 – 674):

“5.5 Impact of the number of clouds on change detection accuracy

         

            The percentage of cloud free pixels appears independent of classification accuracy (Table 9). We therefore find that in the current study sufficient cloud free images were available to not impede or bias the detection of change using the method.  No testing was undertaken to relate the accuracy of the method to the number and temporal distribution of cloud-free observations. This can be addressed in the future by randomly deleting observations and applying the method.

.”

Discussion should discuss a bit about computational aspect of the proposed method as well. I believe it would take quite long to process a large study area (especially when higher resolution imagery is used, e.g., Seninel-2). I also wonder if we use linear trends (or simple mean NDVI) instead of harmonic curves to detect time of change to reduce computational time of this process.

Thank you for this suggestion. Computational complexity has not been a consideration in the work to date, and there has been no attempt at optimization. We have now included computation time for each step in the manuscript. As this is a pixel-based method, we would expect that increasing the resolution or spatial extent, will increase computation proportional to the number of pixels. We do not think the harmonic curves present much different computation time than linear function because they are estimated using a linear regression on cosine transformed NDVI values.

Section 5.3 has been included addressing your comment on lines 645 – 651:

“5.3 Computation time

            Computationally, the most time-consuming step was the linear regression, estimation parameters of equation 2. Note that this step requires multiple linear regressions for each pixel. This was performed on GEE cloud servers and took a maximum of 24 hours. The implementation of equation 3 and the Random Forest classification took less than 10 minutes. As a pixel-based algorithm, the computation time is expected to vary proportionally with the number of pixels.” 

 

Lines 40-46: Incomplete literature review. How about PCC based on a multiple-year dataset (Lark et al.,2015; Nguyen et al., 2019) or a ground sample dataset (Olofsson et al., 2014)?

We have expanded the literature review (lines 47 – 52):

“Some of this error may be mitigated by a process of temporal filtering. This process is applied at the level of pixels and involves the derivation of change trajectories from multi-temporal land cover and some form of correction for implausible sequences of land cover change [7]–[9] . This process is demonstrated to reduce error, but can only do so for restricted types of misclassification and hence can introduces a bias to the analysis.”

And discussed our preference for higher quality ground sample data, but this was unavailable in section 3.1 (lines 145 – 148):

  “Google Earth images represent the highest resolution source of ground truth data available to this study, whilst site visits would likely yield higher quality data, the retrospective nature of this study made this impossible [34]”

 

Lines 71-72: This assumption may not be true for many regions around the world. Why makes this assumption?

Thank you for pointing this out, this has now been deleted

 

Figure 1: texts on the insert map are not clear, remove coordinates of the insert map to avoid confusion.

We have remade the inset map to be clearer. 

 

Line 117: remove “atmospherically corrected”

Thank you for spotting this, it has been removed. 

 

Lines 123-125: do the authors do anything to mitigate this problem? If not, don’t mention it here.

Section 3.2 and 3.3: It would be perfect to couple the ground dataset with the UK’s land cover maps using the protocol presented in Olofsson et al., 2014.

We do not attempt to mitigate this issue and it has therefore been removed. 

We have included two sentences in section 3.2 (lines 145 – 148): 

“Google Earth images represent the highest resolution source of ground truth data available to this study, whilst site visits would likely yield higher quality data, the retrospective nature of this study made this impossible (Olofsson et al., 2014)”

And lines 198 – 200:

“(Olofsson et al., 2014) discusses issues with the manual classification of ground truth pixels.  To address these issues, all the classifications, timings, and justifications are included in the supplementary data (S1 – Points).”

 

 

Line 240: h should be between 0.85 and 1 (as implemented in Figure 4)

Thank you for raising this.  h is a theoretical number that can vary between 0 and 1.  However, we only tested values above 0.85.  We have therefore kept that line the same, and added a clarification to section 4.1 (lines 378 – 379):

“We tested all values of h between 0.85 and 1 in steps of 0.01.”

 

Section 3.4.5: should state all classified classes and their abbreviations here.

Lines 292-293: should put “V – V”, “V – U”, etc... here

Figure 8: set up abbreviations before using them.

Table 4, 8: can use abbreviations to shorten row/column names (e.g., OA, PA, UA, V -V)

Thank you for your final four comments, we have corrected the use of abbreviations Throughout the manuscript.

 

We would finally like to once again thank this reviewer for their time in formulating their response to our manuscript.  We hope that we have made satisfactory improvements to address your comments. 

 

 

 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

This study developed a method that identifies the change in the urban fabric at the pixel level based on breaks in the seasonal and
year-on-year trend of the Normalised Difference Vegetation Index (NDVI).
The method proved relatively robust for outliers and missing data, for
example in the case of high levels of cloud cover, but does rely on a
period of data available before and after the change event.
This method identifies the pixels where land cover changes have occurred with a user accuracy of 45%. This accuracy is too low to apply their method on a large scale. The author needs to give a reasonable explanation for why it is low.

There are many errors in the manuscript, especially obvious reference errors.  Authors should first examine the entire manuscript carefully so that readers can be convinced of their conclusions. In addition, some chapters of the article can be more clear and concise.
Before publishing in Remote Sensing, another round of review should be conducted.

Author Response

Please see attachment. 

Author Response File: Author Response.docx

Reviewer 2 Report

Authors have addressed my concerns during the previous revision, and I recommand for publication. 

Author Response

Please see attachment

Author Response File: Author Response.docx

Reviewer 3 Report

I believe the authors have successfully addressed all my comments. The ms should be ready for publication. Only one final comment for this ms

Lines 49-51: need to elaborate this statement.

Author Response

Please see attachment

Author Response File: Author Response.docx